mCGCNN is a dual-stream crystal graph convolutional neural network for magnetic property prediction. It augments the full crystal graph with a magnetic subgraph that encodes metal–ligand–metal exchange geometry (Goodenough–Kanamori–Anderson rules), then predicts the DFT total magnetic moment per unit cell in μB. Saturation magnetization (Ms / μ₀ Ms) is derived from that moment and the CIF cell volume.
Best for ligand-bridged magnets (oxides, nitrides, and other M–X–M systems). Not recommended for elemental metals or alloys without bridging ligands — those are out of distribution for this checkpoint.
Input structures must contain at least one magnetic site (transition metal, lanthanoid, or actinoid).
Paper: https://arxiv.org/abs/2606.28458 Code: https://github.com/SouravMal/mCGCNN
Is supplement to
Sourav Mal & Satadeep Bhattacharjee · 2026